FeynmanBench: leaderboard

FeynmanBench (2026) gives a multimodal model a machine-generated Feynman diagram (Standard Model QED, electroweak or QCD, tree level to two loops) with minimal labelling conventions and asks it to recover the full physical content of the diagram. Score is the mean of the five checkpoint pass rates, from vertex inventory to the full amplitude, in percent; higher is better.

Metric: Score (%), the mean of the five checkpoint pass rates (vertex inventory, propagator inventory, topological connectivity, momentum routing, full amplitude) on FeynmanBench's Feynman-diagram tasks (a rendered diagram plus minimal conventions; Standard Model QED, electroweak and QCD diagrams from tree level to two loops), API models at temperature 0, mean of five runs; four models ran on subsets of 165 to 742 items, the others on about 2,000; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 17 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)44.3#99
2Gemini 3 Flash40.2#175
3Qwen 3.6 Plus29.6#238
4Seed 2.0 Pro28.6#222
5Claude Sonnet 4.628.5#173
6Qwen 3.5 Plus27.7#242
7GLM-5V Turbo21.3#360
8GPT-5.112.4#240
9GPT-5.412#135
10Claude Opus 4.712#85
11GPT-5.211.1#185
12Claude Opus 4.610.7#113
13Claude Opus 4.510.5#160
14Kimi K2.53.2#266
15Qwen 3 VL Plus2.9#512

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=feynmanbench · How It Works · Data refreshed daily, snapshot 2026-11-21.